7 ms·

every now and then I do something while driving that, IMO, requires "human cognition". Not just image recognition, obstacle avoidance, rule following, etc. Lik

every now and then I do something while driving that, IMO, requires "human cognition". Not just image recognition, obstacle avoidance, rule following, etc. Like, I'm cruising along at 45mph and a plastic bag is blown in front of my car. Yes, I see an obstacle and need to make a decision about what to do - but I do this as a human who knows what a plastic bag is, and how it is far safer to drive right over it than slam on the brakes. I know a plastic bag is not, say, a Pomeranian. Another example: I'm driving, and from beyond my plane of vision, a basketball bounces onto the road up ahead. I can't see past a building, but I know I'm near a park where kids frequently play, and I know it's possible a kid comes running into the road to grab the ball. I proactively slow down. I guess you could say a leavel-5 autonomous car would see the ball and slow down anyway, but there's still that lack of cognition that concerns me. We like to think of driving as a purely rule-based game that is simple enough to model and train against. And I believe in 99% of situations, it is. But in that 1%... the plastic bags and all that...

mgalgs 7y ago
> Not just image recognition, obstacle avoidance, rule following, etc. I would argue that the examples you gave are just image recognition, etc. What is your brain doing if not image recognition and obstacle avoidance? We're not there with today's technology but eventually our machines will perform better at those uniquely human tasks you mentioned. After all, we humans and our cognition is just the output of these giant meat computers inside our skulls pulling in sensory data and running pattern recognition algorithms. The $65,000 question is whether or not today's technology is sufficient (however barbaric) to be acceptable for the job.
- rayiner 7y ago
  > I would argue that the examples you gave are just image recognition, etc. What is your brain doing if not image recognition and obstacle avoidance Planning, anticipation, reading body language, etc? Is that person looking like he’s going to jaywalk? That’s not image recognition. Yes, once he steps into the road the self driving car will respond quicker. But the human driver might already be slowing down and watching for hints of initiating movement. Image recognition is just one tiny piece of it.
  - mgalgs 7y ago
    > That’s not image recognition. How do you do it? Telepathically? You're still just receiving visual input and processing it.
    - rayiner 7y ago
      "Image recognition" is a small subset of "receiving visual input and processing it." We know how to do image recognition with machine learning. Planning and anticipation is a much harder problem.
    - beerandt 7y ago
      Human knowledge and intuition. Deductive reasoning. Knowing there is a location ahead without a crosswalk that kids constantly cross at. Or knowing that the car in front of you is going too fast, implying that they're unaware of the railroad tracks they're about to jump. You don't need telepathy to make useful predictions. And humans don't need to be accurate, they just need to be prepared to take a measured response.
      - Dylan16807 7y ago
        > Knowing there is a location ahead without a crosswalk that kids constantly cross at. That's not based on reasoning. Most people driving by probably don't even know. And it might be in the car's database. > Or knowing that the car in front of you is going too fast, implying that they're unaware of the railroad tracks they're about to jump. If I'm behind them, and I'm going much slower because I know the layout of the road... I'm not sure how knowing this matters?
      - jodrellblank 7y ago
        Simply track all the pedestrians all the time, and you'll know in advance where the common crossing points are, and when any pedestrians are moving towards a road where your car is moving. Apple, Google, Facebook, have every advantage here. Tesla track where every Tesla goes, all the time. Will Apple make a car before Tesla makes a 'smartphone'?
    - deleted 7y ago
- rimliu 7y ago
  No, it is 'model if the world' vs. image recognition. There will be no self-driving cars before the former is available, i. e. for a very long time.
jonplackett 7y ago
I think that sums it up really well. In those kind of edge cases like the basket ball, if an AI driver didn’t slow down and then killed a kid, we’d all be up in arms about It saying how obvious it was that a kid would run out, and we’d have stopped, and we’d be right. But at the same time that AI driver will have stopped in time when 10 other kids jumped out without any warning because it has much, much better reflexes. But it’ll be the first instance that will get all the airtime.
- rubicon33 7y ago
  Nearly 1.25 million people die in road crashes each year. Think about that. Self driving cars aren't just a solution to nuisance of driving. They're a solution to one of the leading causes of DEATH and INJURY in the country. Sure there are going to be edge cases where we as humans will be able to point at the car and say "damnit, we could have prevented that"! But I sure hope when we do that, the car points right back at us and asks "And for the other 1.20 million lives I've saved this year? You would have killed them"
  - rayiner 7y ago
    > Sure there are going to be edge cases where we as humans will be able to point at the car and say "damnit, we could have prevented that"! This is a meaningless assertion unless you account for how often "edge cases" arise. If unpredictable pedestrians, unplanned road construction, etc., is an "edge case" the car basically won't work in DC or New York.
    - smt88 7y ago
      It doesn't need to work in every place to be useful or worthwhile. Sleepy truck drivers are a major safety issue by themselves, and highways are easier for SDCs to drive on.
    - volkk 7y ago
      i would assume that for the first X years, aka a long time, the SDCs would operate mostly on highways/very rural areas.
fwip 7y ago
The driverless cars out there can't even stop for children that are clearly visible. https://www.zdnet.com/google-amp/article/does-your-car-have-automated-emergency-braking-its-a-big-fail-for-pedestrians/ https://www.zdnet.com/google-amp/article/does-your-car-have-... > Using a child-sized dummy, vehicles only avoid running over the child 11% of the time They're not safe.
- jonplackett 7y ago
  This is just for automated braking, not autonomous cars. Not the same thing at all.
  - fwip 7y ago
    The state of the art systems designed only to brake before hitting obstacles can't even do that. It's a pretty simple problem, as far as self-driving tasks go. I don't think that bodes well at all for full autonomy - especially since at least one of those vendors (Tesla) is promising self-driving sooner than later.
  - beerandt 7y ago
    Honest question. Are they fundamentally different, design wise? Or is it merely a measure of confidence & reliability?
    - bryan_w 7y ago
      Self driving cars have more sensors
Dylan16807 7y ago
Even if the car occasionally stops for plastic bags, the self-driving is still a massive improvement in the experience I have. For children you might preemptively slow down, but a properly designed system can brake a whole lot faster. Should be fine.
- beerandt 7y ago
  These are both faulty assumptions. 1) If I'm considering whether to brake for a plastic bag, I'm also going to check for anyone tailgating me, and take that into consideration. Including whether said tailgater is a geo prism or a Freightliner attached to a tanker of gasoline. Slamming on the brakes for a plastic bag is not the universal safer choice. 2) Already mentioned this, but reaction/breaking time is meaningless without timely recognition.
  - Dylan16807 7y ago
    > Slamming on the brakes for a plastic bag is not the universal safer choice. ...I didn't say it was. And you call me the one making assumptions. > reaction/breaking time is meaningless without timely recognition. If it can't quickly recognize large moving objects then it's not qualified to drive at all. I'm assuming it can do that because self-driving at all is part of the premise. (Well, specifically the recognition is necessary for basic driving functionality, and being faster is a matter of adding compute power and using cameras at least as fast as the ones in an iphone.)
    - beerandt 7y ago
      1) I'll rephrase: occasionally (and needlessly) stopping for a plastic bag (or any other false positive) should not be considered a massive improvement over current human abilities 2) >>>you might preemptively slow down, but a properly designed system can brake a whole lot faster. >If it can't quickly recognize large moving objects I'm not questioning whether it can recognize quickly. Just whether it can ever be better/ faster than humans, taking into account cognitive ability. It's whether human (prediction-or-recognition time + reaction time + breaking time) is more or less than (AI recognition time + plus reaction time + breaking time) If AI can react and brake 2 seconds faster than a human, but doesn't recognize the threat until 4 seconds after humans can, then AI drivers will kill more people than human drivers. Which won't be fine. >being faster is a matter of adding compute power and using cameras There are other comments about the limitations of relying only on image recognition. While compute power per might improve the AI recognition time, humans still have the advantage of being able to predict instead of react. This will continue to be a major limitation for AI.
seanmcdirmid 7y ago
How many parks in the USA have speed zones around them that aren’t restricted to 15-25mph? It is stupid easy to slow down for a kid chase ball scenario when you are already going slow anyways, the only reason it is such a problem at all is because human drivers prefer speeding so much.
- rayiner 7y ago
  Tons of them? I don't think any of the parks around here in exurban maryland have any sort of special speed zone. My 100+ year old neighborhood doesn't even have its layout mapped correctly in Google Maps. (The road names in the app don't match the street signs and the wayfinding doesn't realize that the road in front of my house is segmented with barriers so you can't drive all the way down it.)
  - seanmcdirmid 7y ago
    Even in that case, the city default speed limit (when not posted) is usually something low like 25 (Seattle is 20!). Yes, many human drivers ignore that, but self driving cars don’t have that luxury. We really should do this like the Dutch: when something problematic is encountered, engineer and regulate the hell out of it. Not just for self driving cars, but it works for humans to some extent that the laws are actually enforced.
nearbuy 7y ago
The flip side is every now and then a situation arises that requires machine reaction times or attentiveness. Waymo cars have driven over 10 million miles so far without any fatal accidents. That's not nearly enough miles to compare against human drivers, but it'll be interesting to see who does better once they log some more miles.
cf141q5325 7y ago
Its not just plastic bags and fringe error cases. As a driver you know that you sometimes have to break the rules. Say an empty truck is parking on one side of the road, blocking the lane entirely, to overtake him you need to make an illegal maneuver overtaking him or stand there holding up traffic for everyone. Or one step further, you have to make room for an ambulance,. I dont think there is a realistic scenario were all rules to drive in normal traffic can be sufficiently covered.
bobsil1 7y ago
Correct, they need a common sense / world model.
darawk 7y ago
While that's true, the autonomous car can also react so much faster than you can. It may be able to largely trade-off your level of contextual awareness for pure reaction speed in most cases.
- rightbyte 7y ago
  Are you sure? Image recognition or radar does not make self driving cars necessarely faster than a cognitive model and eye input just becouse computers are way faster at math and IO. You need a certain hysteresis/threshold to filter out false positives.

mgalgs 7y ago

> Not just image recognition, obstacle avoidance, rule following, etc. I would argue that the examples you gave are just image recognition, etc. What is your brain doing if not image recognition and obstacle avoidance? We're not there with today's technology but eventually our machines will perform better at those uniquely human tasks you mentioned. After all, we humans and our cognition is just the output of these giant meat computers inside our skulls pulling in sensory data and running pattern recognition algorithms. The $65,000 question is whether or not today's technology is sufficient (however barbaric) to be acceptable for the job.

rayiner 7y ago
> I would argue that the examples you gave are just image recognition, etc. What is your brain doing if not image recognition and obstacle avoidance Planning, anticipation, reading body language, etc? Is that person looking like he’s going to jaywalk? That’s not image recognition. Yes, once he steps into the road the self driving car will respond quicker. But the human driver might already be slowing down and watching for hints of initiating movement. Image recognition is just one tiny piece of it.
- mgalgs 7y ago
  > That’s not image recognition. How do you do it? Telepathically? You're still just receiving visual input and processing it.
  - rayiner 7y ago
    "Image recognition" is a small subset of "receiving visual input and processing it." We know how to do image recognition with machine learning. Planning and anticipation is a much harder problem.
  - beerandt 7y ago
    Human knowledge and intuition. Deductive reasoning. Knowing there is a location ahead without a crosswalk that kids constantly cross at. Or knowing that the car in front of you is going too fast, implying that they're unaware of the railroad tracks they're about to jump. You don't need telepathy to make useful predictions. And humans don't need to be accurate, they just need to be prepared to take a measured response.
    - Dylan16807 7y ago
      > Knowing there is a location ahead without a crosswalk that kids constantly cross at. That's not based on reasoning. Most people driving by probably don't even know. And it might be in the car's database. > Or knowing that the car in front of you is going too fast, implying that they're unaware of the railroad tracks they're about to jump. If I'm behind them, and I'm going much slower because I know the layout of the road... I'm not sure how knowing this matters?
    - jodrellblank 7y ago
      Simply track all the pedestrians all the time, and you'll know in advance where the common crossing points are, and when any pedestrians are moving towards a road where your car is moving. Apple, Google, Facebook, have every advantage here. Tesla track where every Tesla goes, all the time. Will Apple make a car before Tesla makes a 'smartphone'?
  - deleted 7y ago
rimliu 7y ago
No, it is 'model if the world' vs. image recognition. There will be no self-driving cars before the former is available, i. e. for a very long time.

jonplackett 7y ago

I think that sums it up really well. In those kind of edge cases like the basket ball, if an AI driver didn’t slow down and then killed a kid, we’d all be up in arms about It saying how obvious it was that a kid would run out, and we’d have stopped, and we’d be right. But at the same time that AI driver will have stopped in time when 10 other kids jumped out without any warning because it has much, much better reflexes. But it’ll be the first instance that will get all the airtime.

rubicon33 7y ago
Nearly 1.25 million people die in road crashes each year. Think about that. Self driving cars aren't just a solution to nuisance of driving. They're a solution to one of the leading causes of DEATH and INJURY in the country. Sure there are going to be edge cases where we as humans will be able to point at the car and say "damnit, we could have prevented that"! But I sure hope when we do that, the car points right back at us and asks "And for the other 1.20 million lives I've saved this year? You would have killed them"
- rayiner 7y ago
  > Sure there are going to be edge cases where we as humans will be able to point at the car and say "damnit, we could have prevented that"! This is a meaningless assertion unless you account for how often "edge cases" arise. If unpredictable pedestrians, unplanned road construction, etc., is an "edge case" the car basically won't work in DC or New York.
  - smt88 7y ago
    It doesn't need to work in every place to be useful or worthwhile. Sleepy truck drivers are a major safety issue by themselves, and highways are easier for SDCs to drive on.
  - volkk 7y ago
    i would assume that for the first X years, aka a long time, the SDCs would operate mostly on highways/very rural areas.

fwip 7y ago

The driverless cars out there can't even stop for children that are clearly visible. https://www.zdnet.com/google-amp/article/does-your-car-have-automated-emergency-braking-its-a-big-fail-for-pedestrians/ https://www.zdnet.com/google-amp/article/does-your-car-have-... > Using a child-sized dummy, vehicles only avoid running over the child 11% of the time They're not safe.

jonplackett 7y ago
This is just for automated braking, not autonomous cars. Not the same thing at all.
- fwip 7y ago
  The state of the art systems designed only to brake before hitting obstacles can't even do that. It's a pretty simple problem, as far as self-driving tasks go. I don't think that bodes well at all for full autonomy - especially since at least one of those vendors (Tesla) is promising self-driving sooner than later.
- beerandt 7y ago
  Honest question. Are they fundamentally different, design wise? Or is it merely a measure of confidence & reliability?
  - bryan_w 7y ago
    Self driving cars have more sensors

Dylan16807 7y ago

Even if the car occasionally stops for plastic bags, the self-driving is still a massive improvement in the experience I have. For children you might preemptively slow down, but a properly designed system can brake a whole lot faster. Should be fine.

beerandt 7y ago
These are both faulty assumptions. 1) If I'm considering whether to brake for a plastic bag, I'm also going to check for anyone tailgating me, and take that into consideration. Including whether said tailgater is a geo prism or a Freightliner attached to a tanker of gasoline. Slamming on the brakes for a plastic bag is not the universal safer choice. 2) Already mentioned this, but reaction/breaking time is meaningless without timely recognition.
- Dylan16807 7y ago
  > Slamming on the brakes for a plastic bag is not the universal safer choice. ...I didn't say it was. And you call me the one making assumptions. > reaction/breaking time is meaningless without timely recognition. If it can't quickly recognize large moving objects then it's not qualified to drive at all. I'm assuming it can do that because self-driving at all is part of the premise. (Well, specifically the recognition is necessary for basic driving functionality, and being faster is a matter of adding compute power and using cameras at least as fast as the ones in an iphone.)
  - beerandt 7y ago
    1) I'll rephrase: occasionally (and needlessly) stopping for a plastic bag (or any other false positive) should not be considered a massive improvement over current human abilities 2) >>>you might preemptively slow down, but a properly designed system can brake a whole lot faster. >If it can't quickly recognize large moving objects I'm not questioning whether it can recognize quickly. Just whether it can ever be better/ faster than humans, taking into account cognitive ability. It's whether human (prediction-or-recognition time + reaction time + breaking time) is more or less than (AI recognition time + plus reaction time + breaking time) If AI can react and brake 2 seconds faster than a human, but doesn't recognize the threat until 4 seconds after humans can, then AI drivers will kill more people than human drivers. Which won't be fine. >being faster is a matter of adding compute power and using cameras There are other comments about the limitations of relying only on image recognition. While compute power per might improve the AI recognition time, humans still have the advantage of being able to predict instead of react. This will continue to be a major limitation for AI.

seanmcdirmid 7y ago

How many parks in the USA have speed zones around them that aren’t restricted to 15-25mph? It is stupid easy to slow down for a kid chase ball scenario when you are already going slow anyways, the only reason it is such a problem at all is because human drivers prefer speeding so much.

rayiner 7y ago
Tons of them? I don't think any of the parks around here in exurban maryland have any sort of special speed zone. My 100+ year old neighborhood doesn't even have its layout mapped correctly in Google Maps. (The road names in the app don't match the street signs and the wayfinding doesn't realize that the road in front of my house is segmented with barriers so you can't drive all the way down it.)
- seanmcdirmid 7y ago
  Even in that case, the city default speed limit (when not posted) is usually something low like 25 (Seattle is 20!). Yes, many human drivers ignore that, but self driving cars don’t have that luxury. We really should do this like the Dutch: when something problematic is encountered, engineer and regulate the hell out of it. Not just for self driving cars, but it works for humans to some extent that the laws are actually enforced.

nearbuy 7y ago

The flip side is every now and then a situation arises that requires machine reaction times or attentiveness. Waymo cars have driven over 10 million miles so far without any fatal accidents. That's not nearly enough miles to compare against human drivers, but it'll be interesting to see who does better once they log some more miles.

cf141q5325 7y ago

Its not just plastic bags and fringe error cases. As a driver you know that you sometimes have to break the rules. Say an empty truck is parking on one side of the road, blocking the lane entirely, to overtake him you need to make an illegal maneuver overtaking him or stand there holding up traffic for everyone. Or one step further, you have to make room for an ambulance,. I dont think there is a realistic scenario were all rules to drive in normal traffic can be sufficiently covered.

bobsil1 7y ago

Correct, they need a common sense / world model.

darawk 7y ago

While that's true, the autonomous car can also react so much faster than you can. It may be able to largely trade-off your level of contextual awareness for pure reaction speed in most cases.

rightbyte 7y ago
Are you sure? Image recognition or radar does not make self driving cars necessarely faster than a cognitive model and eye input just becouse computers are way faster at math and IO. You need a certain hysteresis/threshold to filter out false positives.